Google Cloud (GPT-6), Anthropic (Duet) and Microsoft (83%) — agents and models in production
Google Cloud puts GPT-6 Astra, Sol and Luna into Model Garden and becomes the only hyperscaler with all three leading closed models. Anthropic opens Claude Duet, a proactive agent on Opus 5.5 with more than 6,000 connectors that works while you are away. Microsoft publishes in Science that MAI-DxO hits 83% on complex NEJM cases versus 36% for general physicians.
The same week any Google Cloud invoice can pay for OpenAI models, an agent starts preparing your work without being asked and an agent orchestrator diagnoses better than the average doctor.
☁️ Google Cloud: GPT-6 on the same invoice as Gemini and Claude
On 21 October OpenAI and Google Cloud announced that GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna enter Model Garden in private preview for enterprise customers. General availability is planned in the coming weeks. Purchases go through Google Cloud Marketplace, draw down existing commitments and appear on a single invoice.
Google thus becomes the only hyperscaler offering the three leading closed models (Gemini, Claude and GPT-6) on the same platform. Microsoft remains OpenAI’s primary partner, but exclusivity is gone. Prompts and outputs stay inside the customer’s data perimeter.
🤖 Anthropic: Claude Duet, the agent that writes to you first
On the same 21 October Anthropic opened the research preview of Claude Duet, a standalone app where Claude acts as a proactive agent. Built on Opus 5.5, it sends a morning brief, works in the background across more than 6,000 connectors and asks for approval before sending messages, paying or signing. Available by waitlist for US Pro and Max subscribers; it shares the plan limits.
It can chase unpaid invoices, prepare payroll or build a dashboard. It never sees passwords directly and a new classifier blocks sharing sensitive data. Anthropic still acknowledges it can misjudge urgency. It is the answer to Meta Muse and OpenAI’s Dots.
🏥 Microsoft: 83% accuracy versus 36% for physicians
On 22 October Microsoft published in Science the results of MAI-DxO, an agent orchestrator that diagnoses sequentially. On 304 complex NEJM cases it reached 83.2% accuracy with GPT-5, versus 36.1% for general physicians with reference access and 22.3% without. On 325 emergency cases it hit 80.3% versus 46.8% for physicians.
The system asks questions, orders tests and debates among agents before committing. In some configurations it cuts estimated cost by half or a third. Microsoft plans real-world tests at Beth Israel Deaconess. Mustafa Suleyman calls it a step toward medical superintelligence. The gain comes from both the base model and the orchestration.
The connecting thread
Frontier models are already billed in the same place the rest of the company’s infrastructure runs. A proactive agent starts deciding what deserves your attention and executing work in the background. Another agent system, in a controlled setting, outperforms the general physician on hard cases. The distance between the lab and production systems is shrinking at the same pace the frictions of permission, cost and liability appear.